Section: Methodology and research methods. Models and forecasts
We assessed the potential of using a computer vision-based neural network to identify undigested parts of the diet in Steller sea lion feces samples. These samples were previously examined and identified by experts. We studied 19 types of bones and otoliths from 13 fish species found in the diet samples and accurately identified by expert morphologists. Each object was photographed against a black background in various projections using a microscope with a +10–15 magnification. This process resulted in 1513 photographs. To identify the undigested diet parts, we utilized the neural network model VGG 16, pre-trained on ImageNet data containing 1.4 million animal and plant images. The model was trained on 1469 photographs of diverse food residues using the R environment and the ‘keras’ package. The training accuracy over 60 epochs reached 99%. We tested the model on 44 images of Steller sea lion diet objects, not used during the training process. The model accurately identified fish remains with 100% accuracy. Computer vision enables quick and precise identification of food residues, reducing the analysis time and cost. It automates the identification process, eliminating human error. Our experiment involved a small data set and needs further research. To improve the feeding objects’ identification accuracy, a larger data set should be used, and the model should be validated on ad ditional test data. The main challenge of using computer vision to identify fish bones is obtaining sufficient photographs of different types of undigested food remains at different stages of digestion from all potential Steller sea lion prey items.
ISSN 1995-4301
Select viewing options
